ZMGA: A ZINB-based multi-modal graph autoencoder enhancing topological consistency in single-cell clustering

The topological structure has consistently been a focal point in single-cell clustering research. Common methods often construct a k-nearest neighbors (KNN) graph from the cell expression matrix, which poses limitations in handling multi-modal data. This issue arises because multi-modal data generate multiple graphs, resulting in inconsistent topological structures. Cells cannot simultaneously maintain contradictory relationships, and this inconsistency may significantly impair the quality of cell representations, ultimately leading to a reduction in algorithmic accuracy. To address these challenges, we introduce a topologically consistent multi-modal graph autoencoder. Specifically, we have developed a Triple-graph Alignment module that utilizes compressed embeddings in the latent space to reconstruct graphs and ensure consistency in the topological structures of the reconstructed graphs with those of each modality. Furthermore, to effectively compress information and accurately model the distribution of real cell data, we have developed both a reconstruction module and a zero-inflated negative binomial (ZINB) module. The reconstruction module restores original information via a compressed hidden layer, thus ensuring efficient information compression. The ZINB module guarantees model conformity to the true distribution of single-cell data. Experiments on six real datasets have validated our model's effectiveness. Our code is publicly available at https://github.com/cywang95/ZMGA.